Stop asking can AI do this. Ask should it.
Can AI do this is almost always yes now, which makes it a boring question. The useful one is whether it should, and what the human keeps. That second question is where all the value lives.
Most people come to AI with the same question. Can it do this thing I do. Can it write the emails, sort the spreadsheet, answer the customers.
It is a fair question. It is also the wrong one, and it has quietly stopped being interesting.
The honest answer to can it is nearly always yes. So if that is the only question you ask, you will hand over far too much, and find out later which parts you should have kept.
Can it stopped being the hard part
A few years ago, can it was a real question with a real chance of no. That is mostly gone. Point a capable model at a task and it will have a go at almost anything, and it will often look convincing while it does.
Which is exactly the trap. Looking convincing and being right enough to trust are not the same thing. Can it answers the first. It tells you nothing about the second.
Should it is the question with the value in it
Should this be done by AI is a completely different question, and a much better one. It asks what you lose by handing it over, not just what you save.
Some work should go. The repetitive, the tedious, the thing nobody enjoys and no customer would notice a machine doing. Hand that over with both hands.
Some work should not, even though it can. The judgement call. The awkward conversation. The thing where being a little bit wrong costs you a lot. Keep those close, on purpose.
There is a real reason should it matters more than can it, and it is in how these tools produce answers. A large language model predicts likely text rather than knowing facts, so it is fluent everywhere and certain nowhere. It will produce a confident answer for the task it should handle and an equally confident one for the task it should not, with no visible difference between the two. That is why you decide the boundary yourself, in advance, rather than trusting the output to flag its own limits. In practice that means setting clear guardrails for what the AI may decide alone, a human checkpoint on anything where being wrong is costly, and a defined standard the output is held against. The model cannot tell you what it should not touch. That judgement is yours to keep.
What does the human keep
This is the part the hype skips straight past. Every sensible use of AI has a quiet second half. The AI does this, and the human keeps that.
What the human keeps is not the leftovers. It is usually the most valuable part. The final read. The relationship. The decision that carries real risk. The taste that makes the work yours rather than generic.
If you only ever ask can it, you never name what the human keeps. So you give it away by accident, and wonder later why the output is competent and somehow not yours.
The people who win are the deliberate ones
I work alongside these tools every day, and the pattern is clear. The people who get the most out of AI are not the ones who hand it the most. They are the ones who were most deliberate about what they handed over and what they held back.
So next time you are tempted by can AI do this, let the answer be yes and move on. The question worth your time is the next one. Should it. And what do I keep.
If you are weighing what to hand to AI and what to keep, that decision is most of the value, and it is exactly the conversation we have before anything gets built.
Book a quick chat →Related: I design systems, not prompts.
Common questions
Should I let AI do this task?
Ask what you lose by handing it over, not just what you save. Repetitive, tedious work that no customer would notice a machine doing is usually safe to give away. Judgement calls, sensitive conversations and anything where being slightly wrong is costly are usually worth keeping, even when AI technically can do them.
Why is can AI do this the wrong question?
Because the answer is now almost always yes, which tells you nothing useful. A capable model will attempt nearly any task and often look convincing doing it. Looking convincing and being right enough to trust are different things, so can it leads you to hand over too much.
What should a human keep when using AI?
Usually the most valuable part: the final read, the customer relationship, the decision that carries real risk, and the taste that makes the work yours rather than generic. Naming what the human keeps is the half of the decision the hype tends to skip.
How do I decide what to automate with AI?
Set the boundary yourself in advance rather than trusting the output to flag its own limits. Give the AI clear guardrails for what it may decide alone, keep a human checkpoint on anything costly to get wrong, and hold every output against a defined standard.